NVIDIA Releases Nemotron 3.5 Transcribe Open Speech Recognition Model
The promptable, open-weight model combines transcription, speaker attribution, translation, and audio understanding into a single mixture-of-experts checkpoint.
In short
- Nemotron 3.5 Transcribe is an open-weight speech model handling transcription across 40 languages, translation across 31, and speaker diarization across 13.
- Built on a hybrid Mamba-Transformer mixture-of-experts design, the model activates roughly 4.3 billion parameters per token out of 30B to 34B total.
- The model placed first among open-weight systems on the Hugging Face Open ASR Leaderboard for English and led speaker attribution on VoiceArena and David AI benchmarks.
- NVIDIA also released the open Granary v2 dataset manifest covering 463,062 hours of speech data alongside NeMo Speech fine-tuning recipes.
NVIDIA on October 28, 2026, released Nemotron 3.5 Transcribe, an open-weight speech recognition model capable of transcribing 40 languages, translating between 31 languages and English, and separating up to eight speakers in 13 languages. Rather than relying on separate pipelines for acoustic decoding, diarization, and text formatting, the model consolidates transcription, speaker attribution, speech translation, and spoken question answering into a single promptable checkpoint under the OpenMDW-1.1 license.
Announced alongside the new Granary v2 dataset manifest, the release targets speech workflows in enterprise meetings, customer support analytics, and medical documentation. The release arrives during an active cycle for open speech technology, following releases like Mistral's Voxtral 3 and Salesforce's Launa-Omni, as documented by TechCrunch.
Architecture and Promptable Capabilities
Traditional speech stacks route audio through specialized speech-to-text encoders, downstream diarization classifiers, and large language models for summarization. According to NVIDIA's Hugging Face model card, Nemotron 3.5 Transcribe operates as a unified Speech-Augmented Language Model (SALM). It pairs a parallel expert audio encoder with a Mamba-2 and Transformer hybrid mixture-of-experts (MoE) backbone.
The full architecture spans 30 billion to 34 billion total parameters, but activates only about 4 billion to 4.3 billion parameters per token. This allows the model to process audio recordings up to one hour long across a 65,536-token context window. To label distinct speakers without post-processing, Sortformer speaker cues feed directly into the language model alongside acoustic features.
Developers guide output through natural language instructions without retraining. A single audio file can yield a verbatim transcript preserving verbal pauses, a cleaned inverse-text-normalized script converting spoken numbers to digits, an English translation, or answers to conversational queries about the recording. For specialized fields, prompts can also include vocabulary hints, such as drug names, which the model incorporates only when acoustic evidence supports them.
NVIDIA emphasized that Nemotron 3.5 Transcribe is an offline processing model designed for recorded audio, not a live streaming speech-to-text service.
Benchmark Performance Across Speech and Diarization
On public evaluations cited in the Hugging Face release notes, Nemotron 3.5 Transcribe achieved top rankings among open-weight architectures.
On the Hugging Face Open ASR Leaderboard, the model placed first among open models in English with a 5.33% average word error rate (WER). In multilingual evaluations across seven languages, it recorded a 4.62% mean WER, and scored 10.98% on long-form audio sets. On the Artificial Analysis Speech-to-Text Leaderboard, it registered a 2.8% AA-WER, placing on the accuracy–active-parameter Pareto frontier by matching the accuracy of 24-billion-parameter systems while using roughly one-sixth of the active parameters. Independent outlet The Decoder noted that the model scored 2.81% in Artificial Analysis's testing, placing slightly behind Mistral's Voxtral Small (2.77%) while requiring far fewer active parameters.
In speaker attribution benchmarks, the model led across multiple evaluations:
- David AI DAI-ASR-I18N: Posted the lowest diarization error rate across 21 languages at 18.26%, edging out NVIDIA's standalone Nemotron 3 Diarization model (18.38%).
- VoiceArena Diarization-Bench: Ranked first among 14 systems in English full-session tests at 32.4% concatenated minimum-permutation WER (cpWER), dropping 8.32% of reference words against a benchmark median deletion rate of 31%.
- Overlap Handling: Accurately maintained speaker tags during rapid, overlapping speech segments, demonstrated in historical multi-speaker film dialogue.
Granary v2 Pipeline and Domain Fine-Tuning
Alongside the model weights, NVIDIA introduced Granary v2, an open data curation pipeline constructed with NVIDIA NeMo Curator. As outlined by SiliconANGLE, the public Granary v2 manifest includes 463,062 hours of transcribed audio across 48 languages drawn from 16 source corpora, expanding from 25 languages in Granary v1. While The Decoder noted total speech assets of roughly 844,000 hours in the broader collection, NVIDIA's official documentation defines the initial manifest release at 463,062 hours.
The pipeline automates the generation of three distinct transcript targets: verbatim text, inverse text normalization with fillers, and cleaned normalized text without disfluencies. It also generates entity hints, question-answering pairs, and code-switching markers across four processing stages: Audio Preparation, Transcription, Task Label Generation, and High Fidelity filtering.
To show how teams can adapt the model to specialized terminology, NVIDIA published an adaptation case study based on Eka Care's Indian medical speech benchmark. Out of the box, standard decoding showed a 30.8% WER. Prompting the model with script and language formatting dropped the error rate to 23.8%. Applying LoRA fine-tuning across the language model and tuning the audio encoder with synthetic speech and clinical samples brought benchmark WER down to 11.8%. On held-out speakers, fine-tuning reduced WER from 22.44% to 12.17%—a 45.8% relative error reduction achieved in 14 GPU hours on a single NVIDIA GB300 system.
Hardware Support and Deployment
Nemotron 3.5 Transcribe weights are available on Hugging Face and ModelScope in full-precision BF16, alongside FP8 and NVFP4 quantized variants.
Serving benchmarks provided by NVIDIA indicate that an RTX PRO 6000 Blackwell Server Edition running NVFP4 precision can transcribe 100 hours of audio in under three minutes at high concurrency. For enterprise deployment, the model can be served using vLLM on a single 80GB H100 GPU. As detailed in the NVIDIA Blog, quantized versions for running Transcribe locally on RTX Spark laptops, DGX Spark systems, and Jetson AGX Thor devices are coming soon, with local C++ runtime support via NeMo-Speech.cpp scheduled for future updates.
Questions people ask
What languages does Nemotron 3.5 Transcribe support?
The model transcribes 40 languages (including seven Indic languages), translates 31 languages to and from English, and performs speaker attribution in 13 languages. Audio question answering is trained in English.
Can Nemotron 3.5 Transcribe be used for real-time live streaming audio?
No. NVIDIA explicitly designed Nemotron 3.5 Transcribe for offline processing of recorded audio up to one hour in length, rather than live audio streams.
What hardware is required to run the model?
The model can be served in full precision on a single 80GB H100 GPU using vLLM. Quantized NVFP4 and FP8 variants run on single-chip platforms including RTX PRO 6000 Blackwell GPUs, while running quantized variants locally on RTX Spark laptops, DGX Spark systems, and Jetson AGX Thor devices is coming soon.
Under what license is Nemotron 3.5 Transcribe released?
The model weights and related artifacts are released under the OpenMDW-1.1 license, permitting commercial and non-commercial development.
Sources
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and AdaptHugging Face, Oct 28, 2026, official source
- Thin, Light and Built for Everything: NVIDIA RTX Spark Windows Laptops Gain New Models, Agent Tools and GamesNVIDIA Blog, Oct 28, 2026, official source
- nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4BHugging Face, Oct 7, 2026, official source
- Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognitionThe Decoder, Oct 28, 2026
- Voice AI in vogue: Nvidia releases new open transcription modelTechCrunch, Oct 28, 2026
- Nvidia releases all-in-one adaptable Nemotron 3.5 audio transcription and translation modelSiliconANGLE, Oct 28, 2026
How this story was made: the newsroom picked it up from huggingface.co, techcrunch.com and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (27 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 29, 2026 at 01:21 UTC
